coding-agent-task-scopinglisted

# Skill: Coding Agent Task Scoping An autonomous coding agent run is not a chat turn. The task description, model tier, and recovery plan must be decided before the run starts — not after it has touched 40 files. This skill structures that pre-run decision. ## Step 0 — Is this task actually agent-appropriate? Not every "I want to automate this" request benefits from an autonomous run. Run this quick check: | Question | If YES | If NO | |---|---|---| | Can the task be done interactively in one supervised chat turn? | → Supervised chat, not an agent run | Continue | | Does the task require reading many files the developer can't enumerate? | → Agent run appropriate | Supervised chat | | Does the task require irreversible production actions (deploys, deletions, DB writes)? | → Mandatory human gate; see high-blast rule | Continue with caution | | Can the task be decomposed into ≤ 5 well-defined sub-steps? | → Prefer sequential supervised turns | Agent run if steps are still complex | ## Step 1 — Define the task boundary A well-scoped agent task has a clear **start state**, **end state**, and **scope envelope**: ``` Start state: [what exists before the run: branch, file list, test state] End state: [what "done" looks like: tests pass, file changed, PR opened] Scope envelope: - Repos in scope: [explicit list] - Files allowed to modify: [pattern or list] - Files NOT to touch: [explicit exclusions — migrations, infra, secrets] - External calls allowed: [none | read-